From mboxrd@z Thu Jan 1 00:00:00 1970 From: Phil Sutter Subject: Re: [RFC net-next] ipv4: Don't promote secondaries when flushing addresses Date: Thu, 7 Jun 2018 14:35:39 +0200 Message-ID: <20180607123539.GH16785@orbyte.nwl.cc> References: <20180607101301.30439-1-jkbs@redhat.com> <20180607110029.vt6tqwlqmuotwpxf@unicorn.suse.cz> <20180607141750.434f6201@beetle> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Michal Kubecek , netdev@vger.kernel.org To: Jakub Sitnicki Return-path: Received: from orbyte.nwl.cc ([151.80.46.58]:33972 "EHLO orbyte.nwl.cc" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753078AbeFGMfo (ORCPT ); Thu, 7 Jun 2018 08:35:44 -0400 Content-Disposition: inline In-Reply-To: <20180607141750.434f6201@beetle> Sender: netdev-owner@vger.kernel.org List-ID: Hi Jakub, On Thu, Jun 07, 2018 at 02:17:50PM +0200, Jakub Sitnicki wrote: > On Thu, 7 Jun 2018 13:00:29 +0200 > Michal Kubecek wrote: > > > On Thu, Jun 07, 2018 at 12:13:01PM +0200, Jakub Sitnicki wrote: > > > Promoting secondary addresses on address removal makes flushing all > > > addresses from a device with 1000's of them slow. This is because we > > > cannot take down the secondary addresses when we are removing the > > > primary one, which would make it faster. > > > > > > However, the userspace, when performing a flush, will in the end remove > > > all the addresses regardless of secondary address promotion taking > > > place. Unfortunately the kernel currently cannot distinguish between a > > > single address removal and a flush of all addresses. > > > > > > To help with this case introduce a IFA_F_FLUSH flag that can be used by > > > userspace to signal that a removal operation is being done because of a > > > flush. When the flag is set, don't bother with secondary address > > > promotion as we expect that secondary addresses will be removed soon as > > > well. > > > > Unless you intend to use the flag to allow deleting a specific address > > with its secondaries (overriding promote_secondaries), maybe it would > > be more practical to go even further and delete all addresses on the > > interface if IFA_F_FLUSH is set so that userspace could delete all > > addresses with one request. > > Thanks for input, Michal. The intend as I understand it is to make > flushing all the addresses fast(er). Let me see if I can rework it > according to your suggestion. It does make more sense to do it like > that to me too. Yes, I agree with Michal. IIRC, flushing a specific primary along with all it's secondaries from an interface is not even supported by iproute2, so no need to optimize for that I guess. OTOH, if your solution allowed to get rid of that nasty loop in ipaddr_flush(), I owe you one extra beer at the next occasion. :) Thanks for holding on to this old ticket! Cheers, Phil